Model Training
Train custom AI models on your specific styles, products, or people for consistent brand visuals.
What is Model Training?
Custom AI: Teach the AI your unique visual style
Consistency: Generate on-brand images every time
Two types: Avatar models (people) and Product models (objects)
Monthly Training Limits
Freelancer: 5 models per month
Agency: 10 models per month
Legacy Plans: Not included
Avatar Models
Preparing Your Dataset
Step 1: Create a Folder
Create a folder with a unique name to avoid confusion with existing AI knowledge:
Good: "Sarah123", "Saravatar", "Saratok", "Emma_tok"
Avoid: Common names like just "Sarah" or "Mike"
Step 2: Add Images
Requirements:
10-20 images of the same person
Same resolution (ideally 1024x1024)
JPG or PNG format
Rename as: img1.jpg, img2.jpg, img3.jpg, etc.
Quality Tips:
Clear, high-resolution photos
Various poses and expressions
Different angles and lighting
Avoid blurry or cropped images
Step 3: Caption Your Dataset (Optional but Recommended)
Using AI for Captions:
Upload images to Claude or ChatGPT
Ask: "Please caption these images. They will be used for a dataset for Lora training."
Save each caption as a text file matching the image name
Example Caption:
"sonya, woman, yellow and black ski jacket, ski goggles, mountain peak background, clouds below, selfie angle, arms extended, snow gear"
File Structure:
YourFolder/ ├── img1.jpg ├── img1.txt ├── img2.jpg ├── img2.txt ├── img3.jpg ├── img3.txt └── ...
Step 4: Create ZIP File
Compress the entire folder into a ZIP file for upload.
Training Your Avatar
Step 1: Upload Dataset
Click the people icon in chat bar
Name your avatar
Upload your ZIP file
Click "Train"
Step 2: Wait for Training
Average time: 5 minutes
You can leave and return
Check status when ready
Step 3: Test Generation
Generate a test image
Always use your avatar name in prompts
Choose:
Approve and save: Deducts 1 training credit
Start over: No credit deduction
Using Your Avatar
For Images
Click the people icon to open modal
Select your avatar from list
Mention avatar name in your prompt
Example: "Sarah123 in a business suit at a conference"
For Videos
Click Studio icon
Select avatar under Assets
Add to scene
Animate or add lipsync
Important Notes
Brand Style Integration:
Avatars inherit your brand's visual style
Same avatar + different brand = different look
Example: 3D illustration brand → 3D-style avatar
Example: Photorealistic brand → realistic avatar
Product Models
Purpose
Show products in different contexts:
E-commerce variations
Marketing materials
Social media content
Package design mockups
Training Process
Prepare products: 15-20 product images
Same format: Consistent resolution and file type
Caption files: Describe each product shot
Create ZIP: Package all files
Upload & train: ~45 minutes processing
Troubleshooting
Quick Checklist
✅ All images in JPG/PNG format
✅ Include captions for better results
✅ Same resolution (728x728 or 1024x1024)
✅ Unique trigger word for avatar
✅ 10-20 images minimum
Common Issues
Poor Results:
Add more variety in dataset
Include captions
Check image quality
Use unique trigger word
Training Fails:
Verify ZIP file structure
Check image formats
Ensure sufficient images
Contact support if needed
Best Practices
Dataset Quality
More photos = better results
Variety improves flexibility
Captions enhance accuracy
Consistency ensures recognition
Prompt Engineering
Always include trigger word
Be specific about context
Add environmental details
Keep requests realistic
Managing Models
Organization
Name models clearly
Document training datasets
Track usage patterns
Delete unused models
Monthly Reset
Limits reset on billing date
Unused trainings don't roll over
Plan training schedule accordingly